Where Static Pipelines Break Down

Most memory-augmented agents still run retrieve-then-reason: retrieval finishes before reasoning starts, and the pipeline can't feed evidence discovered mid-inference back into the retrieval path. A June 2026 paper from researchers at the National University of Singapore, "Memory is Reconstructed, Not Retrieved," argues this rigidity is why agents miss answers on multi-hop questions or drag in irrelevant documents that just inflate the prompt.

A Cue-Tag-Content Graph, Reconstructed on the Fly

The paper's MRAgent stores memory as a three-layer graph — cues, tags, and content — where tags act as semantic bridges linking fine-grained cues to stored content. Instead of finishing retrieval in one pass, it folds LLM reasoning directly into memory access, repeatedly widening or pruning the exploration path based on evidence accumulated so far. The payoff shows up in tokens: per-query prompt consumption fell from 632K for A-Mem and 3.26M for LangMem down to 118K for MRAgent, with accuracy gains over baselines across models and question types.

From Exploration to Stop: A Graph-Memory Rollout Checklist

Set numeric targets before migrating. Cap per-query prompt tokens at 200K — headroom above the paper's 118K observation — and require exploration to converge within 3 hops. Aim for a 15-point-or-better accuracy lift on multi-hop questions over single-pass vector search, and hold p95 latency per reconstruction under 4 seconds.

Lock the tag schema early, too. Over-tagged content nodes make exploration fan out radially and blow through the hop cap; under-tagged nodes never trigger reconstruction and fall back to plain vector search. Keeping roughly 8 content nodes per tag on average is a reasonable starting point.

Failures cluster into three shapes: explosion, where exploration keeps widening with no convergence condition; drift, where stale tags point at outdated content and reconstruction stalls on the wrong node; and idling, where the loop keeps running hops without surfacing new evidence.

Recovery differs by shape. For explosion, cap both hop count and explored-node count, and force an early answer from whatever evidence has accumulated once either limit is hit. For drift, version the tag-to-content mapping and fall back to vector search whenever a stale mapping fires. For idling, stop reconstruction and answer from accumulated evidence once two consecutive hops surface no new node.

Before launch, re-measure both accuracy and token cost on golden sets like LoCoMo and LongMemEval. Mask cue and tag fields before storage so no PII ends up embedded in graph nodes, and log hop count, explored-node count, final token spend, and early-stop reason as required fields — that's what lets you tell reconstruction failures apart from plain vector-search failures.

Roll out in canary, not in one cutover, running alongside the existing vector-only path. Start at 10% of traffic, expand in stages once the token-cost and accuracy deltas hold inside target, and route back to vector search the moment they don't.

Each week, group reconstruction traces that failed on drift or early-stop by type and use them to revise the tag schema, pruning mappings that keep going stale. Keeping a table of which tag clusters have a falling new-evidence-per-hop rate gives the next schema revision a priority list to start from.

Summary

Treating retrieve-then-reason as a fixed, static pipeline means agents miss answers on multi-hop questions or just burn tokens instead of finding them. MRAgent's observation is that combining a cue-tag-content graph with active reconstruction can cut per-query tokens up to 27x against LangMem — but you only get there by capping hop count, versioning tags, and setting early-stop rules first, which is what keeps explosion and drift failures in check.

References

Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — arXiv:2606.06036

New agentic memory framework uses 118K tokens per query, LangMem burns through 3.26M — VentureBeat